A straight-line extractor that produces line descriptions from aerial images is described. The input to the line extractor is in the form of an edge image, where the contrast and direction of each edge pixel is specified. The system scans the edge image left to right and top to bottom and assigns a line label for each scanned edge pixel, thereby generating a label image. At the end of this process, each edge pixel has a line label associated with it, and edge pixels that belong to the same line will be assigned the same line label. In addition, with each line label, a record that stores the end points, the average contrast, and the pixel support of the line is generated. The label image is used as a spatial index to further link fragmented lines. The authors also describe techniques for eliminating many of the physically insignificant lines, given that the domain of interpretation is aerial images dominated by man-made objects.>

A feature based stereo matching system is designed. A hierarchical grouping process that groups line segments into more complex structures that are easier to match is proposed. The hierarchy consists of lines, vertices, edges and surfaces. Matching starts at the highest level of the hierarchy (surfaces) and proceeds to the lowest (lines). Higher level features are easier to match, because they are fewer in number and more distinct in form. These matches then constrain the matches at lower levels. Perceptual and structural relations are used to group matches into islands of certainty. A truth maintenance system (TMS) is used to enforce grouping constraints and eliminates inconsistent match groupings.>

A general framework for building aerial image interpretation systems is proposed. Within this framework, scene knowledge, which is frequently hierarchical and structured in nature, is represented using frames. Domain problem-solving knowledge, which is typically heuristic in form, is declaratively specified through the use of production rules. The issue of efficient search is addressed by the use of an assumption-based truth maintenance system (ATMS). Within these guidelines, a system has been implemented with the goal of detecting a class of buildings in aerial images. The process of building detection is carried out in a hierarchical manner. Line segments are first grouped to form vertices, which are then grouped to form edges. Edges are composed into edge rings. Shadow analysis on closed rings is used to make roof hypotheses. All the elements in the hierarchy are represented by frames. A search scheme based on an ATMS is used to deal with the ambiguities in this hierarchical grouping process.>

The edge-detection problem is posed as one of detecting step discontinuities in the observed correlated image, using directional derivatives estimated with a random field model. Specifically, the method consists of representing the pixels in a local window by a 2-D causal autoregressive (AR) model, whose parameters are adaptively estimated using a recursive least-squares algorithm. The directional derivatives are functions of parameter estimates. An edge is detected if the second derivative in the direction of the estimated maximum gradient is negatively sloped and the first directional derivative and a local estimate of variance satisfy some conditions. Because the ordered edge detector may not detect edges of all orientations well, the image scanned in four different directions, and the union of the four edge images is taken as the final output. The performance of the edge detector is illustrated using synthetic and real images. Comparisons to other edge detectors are given. A linear feature extractor that operates on the edges produced by the AR model is presented.>

An edge detector using the first and second directional derivatives of a random field model is described. The method consists of representing the pixels in a window by a 2 - D causal autoregressive (AR) model, whose parameters are adaptively estimated using a Kalman filter. Due to the modelling assumption, the directional derivates are functions of AR parameter estimates. An edge is detected if the second derivate in the direction of the estimated gradient is negatively sloped, the first derivatives and a local estimate of sample variance are over some threshold values. We illustrate the performance of the edge detector using real image examples.

A new noniterative technique for all pole modelling of noisy speech is proposed. We exploit the additive and high resolution properties of a group delay function to identify the high signal to noise ratio (SNR) regions of the short time spectrum and to separate out these regions from the low SNR regions. A modified spectrum is derived from the group delay function in these regions. The new spectrum is used to derive the all pole model for the speech segment. The main advantage of this method is that it is not necessary to have prior knowledge of the noise or speech characteristics as in the other methods of processing noisy speech.